my friends have you heard the good news about our lord and savior posits?
|
|
# ? Oct 21, 2023 06:51 |
|
|
# ? May 26, 2024 20:35 |
|
rjmccall posted:cross-posting from the buttcoin thread: respect to the Thorne toe
|
# ? Oct 21, 2023 07:23 |
|
VikingofRock posted:my friends have you heard the good news about our lord and savior posits? oh is that dude still trying to push his bullshit?
|
# ? Oct 21, 2023 08:07 |
|
it is a good time for being on some bullshit in the area, as the ieee-754 hegemony is a fair bit weaker than it has been in a number of decades.
|
# ? Oct 21, 2023 11:05 |
|
redleader posted:floating point is a mistake and anyone who says otherwise is gaslighting you floating point is great, actually
|
# ? Oct 21, 2023 14:03 |
|
DELETE CASCADE posted:good for video games tho Video games would have been fine with massively parallel fixed point coprocessors. The thing that matters for games is pushing lots of operations, not the specifics of float which are frequently inconvenient.
|
# ? Oct 21, 2023 14:10 |
|
Cybernetic Vermin posted:it is a good time for being on some bullshit in the area, as the ieee-754 hegemony is a fair bit weaker than it has been in a number of decades. in what way is it weaker? the most popular "non-ieee" format i've heard of is bfloat16, and i put scarequotes on that because it's literally 32-bit ieee-754 with fewer mantissa and exponent bits the only real non-754 motion i'm aware of (and by "real" i mean actually generating hardware implementations) is that some ML applications want 8-bit (or even smaller) floats and 754-style floats have problems scaling down to so few bits i don't think that's going to scale back up and take anything over though, it's just ML folks being willing to live with a very tailored and limited numeric format which doesn't make much sense outside their problem domain the unum / posit guy is a crank who's intensely jealous of william kahan, the guy who led the ieee-754 effort. unfortunately for his crusade to discredit kahan and put himself on the floating point throne, posits are shite BobHoward fucked around with this message at 00:06 on Oct 23, 2023 |
# ? Oct 23, 2023 00:03 |
|
BobHoward posted:in what way is it weaker? the most popular "non-ieee" format i've heard of is bfloat16, and i put scarequotes on that because it's literally 32-bit ieee-754 with fewer mantissa and exponent bits just fewer mantissa bits, the whole point is it’s a much less precise float and has the same exponent range
|
# ? Oct 23, 2023 00:10 |
|
implement arbitrary precision math in silicon or get the hell out of my fab
|
# ? Oct 23, 2023 06:12 |
|
redleader posted:implement arbitrary precision math in silicon or get the hell out of my fab god gave you ADC and SBB
|
# ? Oct 23, 2023 06:41 |
|
leper khan posted:Video games would have been fine with massively parallel fixed point coprocessors. The thing that matters for games is pushing lots of operations, not the specifics of float which are frequently inconvenient. Urgh no thanks i worked on playstation 1 and nintendo ds games and fixed point was a pita I prefer floats, pitfalls warts and all
|
# ? Oct 23, 2023 11:38 |
|
Yeah, the only folks who think simulating a game world with fixed point makes anything easier are those who haven't tried. Not needing to separately reason about the useful range and potentially need to adjust the representation of every single intermediate result is actually amazing.
|
# ? Oct 23, 2023 18:26 |
|
see, like, that's the thing about floating point, it's better than what people try to replace it with for ex: people get mad about negative zero, but that's because they think of floating point as "exact numbers" over "approximate numbers after rounding". in reality, -0 is "a very small negative number that's effectively zero" much in the same way +0 is "a very small number that's effectively zero", and that's why they're different. sign preservation is useful people get mad about NaN, and i am sympathetic, but there is a real use. if you're doing a big rear end calculation, you don't want to do error checking after every operation. it makes it hard to do pipelining or superscalar stuff or whatever the kids are into these days, but in more practical terms: error checking after every arithmetic op greatly inflates the size of your program and sabotages the speed. therefore, floating point has an error value, NaN, and just to ensure that "f(x) == f(y)" doesn't accidentally do the right thing, let's make it so one error can never equal another. then comes subnormal and gradual underflow. if you ask me, this is the neatest trick that floating point handles. when you're storing a number as a float, you could represent the same number in different ways, like 10 as 100 ** -1 or 10 ** 1, and so floating point specifies that all numbers have a singular standard form. floating point then goes onto say "actually, here are some not normal ones, just so we can extend the range of precision". it's wringing out all the last drops of precision and it's ingenious if there's any real improvement to be made for floating point, it's to have a decimal coded mantissa. not for any real accuracy or precision, but because it'll be a little more humane when 0.1+0.2 == 0.3
|
# ? Oct 23, 2023 19:22 |
|
numbers are just cursed, floating point's just a little bit more obviously so
|
# ? Oct 23, 2023 19:22 |
|
i remember java3d had this notion of arbitary-precision origins that let you just use regular floats associated with that origin and retain as much resolution as you liked to because you could provision an origin whereever you wanted one, and you only paid a big price on operations when you had to do math with objects in different origin spaces and that always seemed like it'd be a good solution for somebodies problem.
|
# ? Oct 23, 2023 19:31 |
|
idk if that was a common solution to things or what but it always seemed neat to me.
|
# ? Oct 23, 2023 19:32 |
|
Ralith posted:Yeah, the only folks who think simulating a game world with fixed point makes anything easier are those who haven't tried. Not needing to separately reason about the useful range and potentially need to adjust the representation of every single intermediate result is actually amazing. Depends what problems you're solving for.
|
# ? Oct 23, 2023 19:42 |
|
tef posted:if there's any real improvement to be made for floating point, it's to have a decimal coded mantissa. not for any real accuracy or precision, but because it'll be a little more humane when 0.1+0.2 == 0.3 herman hollerith's gang has your back, they got a standard for decimal floating point into ieee 754-2008 you have to buy ibm mainframe hardware and operating systems to actually use it, but it's out there
|
# ? Oct 23, 2023 20:57 |
|
Adding a second sign bit so I can do complex arithmetic more easily. Four zeroes
|
# ? Oct 23, 2023 21:24 |
|
DB/2 does has a decimal type that uses the POWER instructions where available, and IBM had the Rexx guy (Mike Cowlishaw, I think) trying to jam decimal math into JavaScript for more than a decade, IIRC. they really like their decimal over there
|
# ? Oct 23, 2023 21:35 |
|
decimal, from the roots deci- meaning "base 10" and -mal meaning "bad" or "wrong"
|
# ? Oct 23, 2023 21:37 |
I just wish floats were associative, and that it was a little easier to do stuff like summing products without losing precision. For all that I was joking about posits earlier, I actually do think the general goals of unums / posits are pretty admirable. I don't really have strong opinions on floating infinities, subnormals, NaN, or negative zero, although I do wish some of the comparison operators involving them were a little less tricky. I once had a hell of a bug involving negative zero though, where a serialization library assumed that if a == b (and they were both floats), then their serializations would also be equal. But it serialized negative zero differently than positive zero (thus violating its own internal assumptions), and it had some very weird behavior because of that.
|
|
# ? Oct 23, 2023 21:40 |
|
NaNs are great because they let you smuggle data in floating point values, and that is a very fun thing to do
|
# ? Oct 23, 2023 21:42 |
|
NaN packing blew my mind when I first learned about it.
|
# ? Oct 23, 2023 21:48 |
|
i just think floating point should be one of those things you opt into, instead of being e.g. the only number type available in the world's most popular plang (bigint doesn't count)
|
# ? Oct 23, 2023 22:49 |
|
What should the default number type be?
|
# ? Oct 23, 2023 22:53 |
|
Athas posted:What should the default number type be? going up
|
# ? Oct 23, 2023 23:05 |
|
redleader posted:i just think floating point should be one of those things you opt into, instead of being e.g. the only number type available in the world's most popular plang (bigint doesn't count) plenty of languages let you pick your number type. even ruby of all languages has separate Integer and Float types and even eventually gained a Decimal type to use instead of Float if you want. does bigint not count since it’s only integers?
|
# ? Oct 23, 2023 23:16 |
|
rotor posted:decimal, from the roots deci- meaning "base 10" and -mal meaning "bad" or "wrong" there's like one, maybe two good usecases for decimal outside of just "humans expect math to work this way" like money one of them is "with binary coded decimal, it's way easier to translate game scores into tile offsets", at least on dmg era gameboys and the other is pretty much the same deal, like facebook found that itoa was taking up a disproportionate amount of time in logging
|
# ? Oct 23, 2023 23:20 |
|
Armitag3 posted:going up "int NOT NULL AUTO_INCREMENT" it is, then!
|
# ? Oct 23, 2023 23:22 |
|
Athas posted:What should the default number type be? bool.
|
# ? Oct 23, 2023 23:26 |
|
there's this long post about posits being a scam https://marc-b-reynolds.github.io/math/2019/02/06/Posit1.html anyway somewhere halfway through it, there's this graph of relative error and posits do not look great i'd attach it but i'm too lazy to resize a screenshot to make it upload
|
# ? Oct 23, 2023 23:27 |
|
redleader posted:i just think floating point should be one of those things you opt into, instead of being e.g. the only number type available in the world's most popular plang (bigint doesn't count) you can opt out of using a lovely plang
|
# ? Oct 23, 2023 23:36 |
|
no i can't, it runs on every loving webpage i visit
|
# ? Oct 23, 2023 23:58 |
|
Athas posted:What should the default number type be? s-expression
|
# ? Oct 24, 2023 00:07 |
|
Athas posted:What should the default number type be? string imo
|
# ? Oct 24, 2023 00:08 |
|
binary coded decimal
|
# ? Oct 24, 2023 00:10 |
|
Athas posted:What should the default number type be?
|
# ? Oct 24, 2023 00:11 |
Athas posted:What should the default number type be? a symbolic expression, which is then computed to arbitrary precision when it's time to output the results
|
|
# ? Oct 24, 2023 00:29 |
|
|
# ? May 26, 2024 20:35 |
|
DELETE CASCADE posted:no i can't, it runs on every loving webpage i visit you can disable javascript if it really bothers you
|
# ? Oct 24, 2023 00:33 |